Transcript Slide 1

Continuation of Lecture 1
Corrections/Clarifications from Day 1
1. Lin-Rood scheme stability comes about because it is equivalent to a
Semi-Lagrangian scheme not using implicit differencing. (Thanks to
Ravi for pointing this out).
2. Definition: An anomaly is the deviation from some mean (frequently
also called a climatology in meteorology/oceanography)
3. What do we mean by seasonal forecasts?
a. Forecast for a three month average of field of interest
(precipitation). For example, forecast for June-July-August (JJA)
starting from May 1.
4. Limit of predictive skill for seasonal forecasts (Tropical Pacific SST)
that I gave (order of six months) is based on skill level that is
realized currently. Here a usual metric for skillful would be an
anomaly correction of 0.6.
Lorenz limit for weather predictability of two weeks is from
perturbation experiments, i.e. how small differences in initial state
grow in time. Currently realized forecast skill is (significantly) less
than two weeks.
Corrections/Clarifications from Day 1 (continued)
4. Continued: You can do same error growth type of calculation for
monthly means of tropical Pacific SST indices (Nino 3). This has
been done with “Intermedate” coupled models such as Cane-Zebiak
model. “Potential” predictability limit from those sorts of equations is
36 months. Currently realized forecast skill is less than this.
5. Based on my comment that the atmospheric initial condition memory
is gone in two weeks then what use is assimilation to longer term
forecasting problem?
a. Memory of the initial state in the ocean (large thermal inertia) is
on the order of at least 6 months for fields of relevance to the
seasonal forecasting problem (Near equatorial mass anomalies
associated with equatorial Rossby and Kelvin waves). Therefore
initial state specification in ocean is very important. The assimilation
also serves another role in correcting the model bias, especially the
structure of the thermocline.
Corrections/Clarifications from Day 1 (continued)
5b. For the decadal forecasting problem different SST anomalies
(modes of variability) are important. Initializing the ocean is
important to capture these modes of variability. This is an area of
active research. Next round of IPCC/CMIP will include extensive set
of decadal forecasts by many international groups.
6. What do skill maps look like for current CGCM SST forecasts?
Maps from CFS will be shown.
7. If you have questions in next few weeks or in the future please send
me e-mail: [email protected]. Please put in the e-mail
Subject: TIFR Summer School.
Seasonal Forecast Skill of SST
1. Forecast skill of is dependant on the initial condition (IC) month. In
the central and eastern Pacfiic the SST forecast skill undergoes
quick dcline in the Northern Hemisphere Spring which is
consequently known as the Spring Predictability barrier.
Examine SST forecast skill from NCEP CFS or 2 IC months: January
and August. Deterministic skill score (ACC).
Define a signal to noise ratio. Signal is ensemble mean standard
deviation and noise is deviation of ensemble members around
ensemble mean.
CFS SST January IC (Anomaly Correlation) (W. Wang)
CFS SST January IC (Anomaly Correlation) (W. Wang)
CFS Signal to Noise Ratio for Jul IC (W. Wang)
Decadal Forecasting
1. Two modes of variability that people hope to predict:
Pacific Decadal Oscillation (PDO)
Atlantic Meridional Oveturning Circulation (AMOC)
2. What do we mean by decadal variability? Interannual variability
around the decadal variability?
3. What are challenges? Prescribing ocean initial state. Difficult do to
Lack of data.
4. Decadal forecast data will be available for anyone to examine and
come to your own conclusions. IPCC AR5.
North Atlantic Temperature
Atlantic Meridional
Overturning Circulation
(AMOC)
Warm North Atlantic
linked to …
Drought
More strong hurricanes
More rain over Sahel
and western India
Two important aspects:
a. Decadal-multidecadal fluctuations
b. Long-term trend
(Courtesy: Joe Tribbia, NCAR)
PDV (PDO): Pacific Decadal Variability
The principal “mode” in the Pacific
POSITIVE Phase
NEGATIVE Phase
PDO refers mainly to N.
Pacific sea surface
temperatures (SSTs).
Climate teleconnections
similar to those of ENSO
Precipitation Correlation
Substantial interannual
variability
Several processes at work
Meehl & Hu, J. Climate, 2006
(Source: http://jisao.washington.edu/pdo)
Decadal Prediction
But there are challenges …
• Initialization
o Many different global reanalysis products, but
significant differences exist
Large inherent uncertainty in driving of AMO
Atlantic Salinity Anomalies (upper 300 m)
12m-rm seasonal anom: NPAC Averaged salinity over the top 300m
12m-rm seasonal anom: TRATL Averaged salinity over the top 300m
0.15
0.6
ukdp
ukoi
cfcs2
cfas2
ecco50y
0.4
gfdl
soda
ecmfa
ecmfc
ukgs
ingv
mrieccoSIO
cfasa
mct2
Tropics

mct3
eccoJPLa
eccoJPLc
eccoMIT
GMAO
0.10
ukdp
ukoi
cfcs2
cfas2
ecco50y
gfdl
soda
ecmfa
ecmfc
ukgs
ingv
mrieccoSIO
cfasa
mct2
mct3
eccoJPLa
eccoJPLc
eccoMIT
GMAO
0.05
0.2
Mid-Lat

0.00
0.0
-0.05
sdv ensm = 0.014
s/n ensm = 0.246
-0.2
1950
sdv all = 0.051
s/n all = 0.912
spread
1970
1980
(Courtesy:1960
Joe Tribbia,
NCAR)
Time
= 0.056
1990
2000
sdv ensm = 0.010
s/n ensm = 0.423
-0.10
1950
1960
sdv all = 0.024
s/n all = 1.015
1970
spread
1980
= 0.023
1990
2000
Decadal Prediction
But there are challenges …
• Initialization
o Many different global reanalysis products, but
significant differences exist
Ocean observing not yet global or comprehensive
Tropical Upper Ocean T Anomalies (Upper 300 m)
12m-rm seasonal anom: EQPAC Averaged temperature over the top 300m
12m-rm seasonal anom: EQIND Averaged temperature over the top 300m
1.5
1.0
0.5
ukdp
ukoi
cfcs2
cfas2
ecco50y
gfdl
soda
ecmfa
ecmfc
ukgs
ingv
mrieccoSIO
cfasa
mct2
Pacific

mct3
eccoJPLa
eccoJPLc
eccoMIT
GMAO
ukdp
ukoi
cfcs2
cfas2
ecco50y
gfdl
soda
ecmfa
ecmfc
ukgs
ingv
mrieccoSIO
cfasa
mct2
mct3
eccoJPLa
eccoJPLc
eccoMIT
GMAO
0.0
0.5
-0.5
Indian

0.0
-0.5
sdv ensm = 0.272
s/n ensm = 1.139
sdv all = 0.337
s/n all = 1.411
spread
sdv ensm = 0.136
s/n ensm = 0.619
sdv all = 0.220
s/n all = 0.998
spread
= 0.220
= 0.239
-1.0
1950
1960
1970
1980
(Courtesy:
Joe Tribbia,
NCAR)
Time
-1.0
1990
2000
-1.5
1950
1960
1970
1980
Time
1990
2000
Progress with imperatives (CLIVAR JSC31)
Decadal variability and predictability
• Some key questions
− To what extent is decadal variability in the oceans
and atmosphere predictable?
− What are the mechanisms of variability?
− Does oceanic variability have atmospheric consequences?
− Do we have the proper tools to realize the predictability?
Need for (coupled) data assimilation systems to initialize models
Are models “good enough” to make skillful predictions?
Adequacy of climate observing system?
Global number of temperature
observations per month as a
function of depth
1980-2006
Timescales of Variability in Observations
e.g. Climate Variability & Change in CO
Temperature
Precipitation
25%
1%
13%
25%
62%
74%
What value is there in interannual signal of long-term forecast?
Suppose we have these decadal forecasts which will have interannual
variability.
The value of the interannual variability part of the forecast would not
be forecasting for a specific year several years in the future. Value
would be in characterizing the statistics of the variability over
some multi-decadal period.
1. What is the probability of JFM rain that exceed some threshold?
2. What is the probability of increases in extreme events of heat,
cold, rain?
This type of information would be useful to infrastructure planning.
Will we be able to do this? No one knows…but this is the type of
information governments are now asking scientists to produce.
Coordinated Decadal Prediction for
AR5
Basic model runs:
1.1) 10 year integrations with initial dates towards the end of 1960,
1965, 1970, 1975, 1980, 1985, 1990, 1995 and 2000 and 2005 (see
below).
- Ensemble size of 3, optionally increased to O(10)
- Ocean initial conditions should be in some way representative of the
observed anomalies or full fields for the start date.
- Land, sea-ice and atmosphere initial conditions left to the discretion
of each group.
1.2) Extend integrations with initial dates near the end of 1960, 1980
and 2005 to 30 yrs.
- Each start date to use a 3 member ensemble, optionally increased to
O(10)
- Ocean initial conditions represent the observed anomalies or full
fields.
Current and Historical State of the Ocean Observing System
1. SST: First satellite based products (global coverage) starting in
1982. So, only about 30 years of data. Clever people have
constructed EOF (SVD) reconstruction techniques to go back
earlier in time when only had ship data. How well do the two
types of products compare for the common era?
2. Sub-surface:
TAO/TRITON Array in Pacific (1990-present (sometimes))*
RAMA: Indian Ocean (
PIRATA: Atlantic
ARGO:Everywhere (starting around 2000 and filling in till now)
• Why sometimes? (Wear and tear)
Buoys in “cold tongue” are good surface for phytoplankton.
Small fish eat the phytoplankton. Big fish eat the small fish.
Fisherman come to catch the big fish and do all sorts of
interesting things to the buoys. None of which are good.
Consistency of Observed Data Sets
O. Alves
Why is TAO/TRITON Array Arranged that way
1. Buoys are expensive so want a minimal set that can do the job.
2. The delayed oscillator theory and importance of equatorial waves
was used to establish the need for the array.
3. Ocean currents are meridionaly confined.
Equatorial Pacific Temperature Anomaly
TAO
GODAS
TAO climatology used
- Note the differences between GODAS and
TAO temperature are as large as 2-3C in
the eastern equatorial Pacific near the
thermocline since mid Jan 2010.
- The large departures from observations
might be related to the failure of the three
eastern most equatorial buoys
(http://tao.noaa.gov).
25
TAO/TRITON Observing Status in July
- TAO moorings had
massive failures at
95W and 110W near
the equator.
http://tao.noaa.gov/tao/status/
26
Equatorial Pacific Temperature
TAO Temp Anom GODAS-TAO
http://tao.noaa.gov/tao/status/
http://www.coriolis.eu.org/cdc
GODAS-Coriolis
- Equatorial temperature
decreased at the surface
and near the thermocline,
probably forced by easterly
wind anomalies in the
central-eastern Pacific
(slide 14).
- Temperature differences
between GODAS and TAO,
GODAS and Coriolis, were
above 1C near the
thermocline east of 135W.
- Those positive biases
were consistent with warm
biases in the control
simulation in which
observations were not
included.
- Therefore, TAO mooring
data at 95W and 110W
played critical roles in
constraining model biases.
27
The Future is Much Brighter for Ocean Observations
ARGO array: Automatic buoys that dive to 1000 meters then
back to surface. Periodically at the surface they transmit
their data to satellite.
Nearly global coverage of the worlds oceans.
Measure temperature and salinity.
Multi-national: Many countries bought buoys.
28
A. Weigel
Indian Ocean
Temperature
Salinity
PreArgo
Argo
O. Alves
The Centre for Australian Weather and Climate Research
A partnership between CSIRO and the Bureau of Meteorology
Example of ODA Estimates of Observed State
In data sparse regions (everywhere outside the Tropical Pacific) and for
fields that are not observed solutions from ODA products can show
substantial variability. This makes their use as a verification product
for models more ambiguous.
Different Realities Simulated by ODA Systems
• Comparison of 20
year climatology of
January surface zonal
current from 3 state of
the art ODA systems:
Rather large
differences especially
near the equator.
Coupled Model Bias (Systematic Error)
1.
SST errors are frequently same order of magnitude as the
Seasonal variability (standard deviation).
2. SST errors are similar across models.
3. Equatorial SST variability has wrong spatial structure and
amplitude.
4. Precipitation structure becomes distorted and too symmetric
about the equator: “Double ITCZ Syndrome”
Use of Multi-Model Ensembles (MME) in Seasonal Forecasting
1.
Consists of different system each with its own ensemble
members.
2.
Has been found that taking more than one forecast system
and combining with other systems produces more skillful
forecasts.
3.
Has also been shown (fewer comparisons) that using the
same number of ensemble members from the best model as
for the MME still has MME with higher skill scores.
4.
Most MME systems are equal weight, i.e. no difference in
weights for good versus bad models. Reason for this is that
with 30 years of data and cross validated weights it is hard to
beat equal weights.
RPSS: temp
1-tier
2-tier
RPSS: pcp
1-tier
2-tier
MME SST is Generally Better Than Any Model
Intercomparison of GCM precipitation seasonal forecast skill
Anomaly correlation
NCEP CFS
JAMSTEC SINTEX-F
ECHAM4.5-MOM3 (DC2)
ECHAM4.5-MOM3 (AC1)
June–Sept
Seasonal total
from May 1
(1982–08)
ECHAM4.5-caSST
ECHAM4.5-EC3_SST
ECHAM4.5-GMLcfsSST
Seasonal Forecasting: Marginal Skill Problem
Show skill scores for actual forecasts over the last 11 years from
IRI. 2-tiered (uncoupled system). Order of 7 models most with
about 10 ensemble members. Forecasts are terciles.
Skill metrics used:
1.
Ranked Probability Skill Score (RPSS)
2.
Liklihood score.
3.
Generalized Relative Operating Characteristics (ROC)
Verification
measure
comparison:
All-season
temperature,
0.5-month
lead time
Verification
measure
comparison:
All-season
precipitation,
0.5-month
lead time
Use of Forecast System Including ODA to evaluate importance of
observation networks
Observing System Simulation Experiments (OSSE):
Run ODA with full data stream and then cases removing select data
types.
Run forecasts from each of these cases and examine difference in
forecasts, especially skill.
Impact of Ocean Observations
O. Alves